ABSTRACT Joint clustering of hyperspectral and Light Detection and Ranging (LiDAR) data is challenging due to their heterogeneity and differing spatial-spectral characteristics. To address this, we propose an adaptive multi-view graph convolutional network (MVGCN) that integrates visual Bidirectional Encoder Representations from Transformers (VisualBERT), referred to as MVGCN-VisualBERT, to extract high-level semantic features from both modalities. These features form a superpixel-level graph that preserves spatial structure while reducing redundancy. A multi-view graph convolutional network then propagates and aggregates information to enhance cluster cohesion. Evaluated on the MUUFL and UH2013 datasets, MVGCN-VisualBERT outperforms state-of-the-art methods, achieving improvements of 2.8% in overall accuracy, 2.6% in the Kappa coefficient, 2.9% in normalized mutual information and 5.8% in the adjusted Rand index on MUUFL. These results highlight the potential of the proposed approach for improving unsupervised multimodal land-cover analysis in remote sensing applications.

Multiscale multimodal graph convolutional networks for identifying essential tremor and dystonic tremor.

Message passing is a core mechanism in Graph Neural Networks (GNNs), enabling the iterative update of node embeddings by aggregating information from neighboring nodes. Graph Convolutional Networks (GCNs) exemplify this approach by adapting convolutional operations for graph structures, allowing features from adjacent nodes to be combined effectively. However, GCNs encounter challenges with complex or dynamic data. Capturing long-range dependencies often requires deeper layers, which not only increase computational costs but also lead to over-smoothing, where node embeddings become indistinguishable. To overcome these challenges, reservoir computing has been integrated into GNNs, leveraging iterative message-passing dynamics for stable information propagation without extensive parameter tuning. Despite its promise, existing reservoir-based models lack structured convolutional mechanisms, limiting their ability to accurately aggregate multi-hop neighborhood information. To address these limitations, we propose RGC-Net (Reservoir-based Graph Convolutional Network), which integrates reservoir dynamics with structured graph convolution. Key contributions include: (i) a reimagined convolutional framework with fixed-random reservoir weights and a leaky integrator to enhance feature retention; (ii) a robust, adaptable model for graph classification; and (iii) an RGC-Net-powered transformer for graph generation with application to dynamic brain connectivity. Extensive experiments show RGC-Net achieves state-of-the-art performance in classification and generative tasks, including brain graph evolution, with faster convergence and mitigated over-smoothing.

CVAE-guided triage and modular classifiers for multimodal ASD detection.

FocusLG: Focusing on local and global molecular representation for kinase inhibitor binding affinity prediction.

Compaction state value: a new indicator for subgrade compaction quality assessment with graph convolutional neural network

DMVHP-IBS: Dynamic feature-integrated multi-modal prediction of virus-host protein interactions and the binding sites.

Dynamic spatial graph convolution and attention network with dual threshold inference for retinal vessel segmentation

Fault prediction for aluminum electrolysis based on spatial-temporal attention dynamic graph convolutional networks

MGEGR: Multi-view graph convolutional networks for ephemeral group recommendation

Tucker decompositions and graph convolution network based radio frequency fingerprint identification with extremely small sample size

Dynamic trust evaluation based on heterogeneous graph convolution network and lightweight attention

SPD-Net: A semantic partitioned transformer with dynamic graph network for improved skeleton-based gait recognition.

A dynamic multi-dimensional globally aware graph convolutional network for traffic prediction

The structural and functional connectivity of the brain network is a combination of complex connections and interconnections among neurons of different brain regions. Analysis of these connectivity patterns provides a representation of the mental health of an individual, including major depressive disorder (MDD). MDD analysis with Electroencephalogram (EEG) has increasingly focused on characterizing alterations in brain connectivity patterns. The dynamic complexities in the functional connectivity of the brain and its intricate interconnections have posed significant challenges in effectively utilizing EEG data and machine learning techniques to extract meaningful information for accurate MDD diagnosis. This paper proposes the use of a graph convolutional neural network (GCNN) for MDD classification and analysis while discovering the dynamic functional connectivity involved through scalp EEG recordings. The connectivity graph was obtained using spectral coherence of resting-state and task-based EEG recordings. This paper makes three key contributions: it proposes a novel GCNN model with a unique graph representation for EEG data that integrates domain knowledge of brain regions with data-driven functional relationships without relying on explicit spatial geometry; it conducts extensive evaluation using two distinct scalp- EEG datasets and identifies effective frequency bands and brain region markers associated with MDD. The proposed method demonstrates strong performance in classifying MDD patients and healthy individuals, achieving an AUROC of 93% for dataset 1 and 71% for dataset 2. The study found consistent identification of neurophysiologically relevant biomarkers across both datasets, including dominant activity in the left prefrontal, left frontoparietal and right prefrontal regions, along with contributions from delta, theta, and alpha frequency bands. In contrast, the lower beta band has only a minimal influence compared to the others.

Circular RNAs (circRNAs) are an emerging class of non-coding RNAs with covalently closed loop structures and have been increasingly recognized for their regulatory roles in disease progression and drug response. Accurately identifying circRNA-drug sensitivity associations is therefore essential for understanding therapeutic mechanisms and advancing precision medicine. However, most existing computational methods fail to effectively integrate semantic and structural information and overlook cross-modal feature co-optimization, thereby limiting their predictive performance. To address these limitations, we develop an end-to-end graph representation learning framework for circRNA-drug sensitivity prediction by jointly modeling homogeneous similarity structures and heterogeneous interaction relationships. The framework integrates fused similarity graphs, semantic feature encoding with pre-norm residual attention, and structural representation learning via graph convolutional networks with Top-K sparse adjacency. In addition, a large-scale heterogeneous graph and a cross-modal collaborative feature mining module are employed to jointly optimize multi-source representations. Experimental results from 5-fold and 10-fold cross-validation, independent test evaluations, ablation study, and case study demonstrate that the proposed framework consistently achieves superior performance compared with state-of-the-art methods. The proposed framework provides a robust and effective computational strategy for circRNA-drug sensitivity prediction and offers a valuable tool for uncovering potential therapeutic associations, thereby facilitating future research in drug response analysis and precision medicine.

RhlA serves as the crucial rate-limiting enzyme in rhamnolipid biosynthesis. In this study, we developed a fusion model named EGCA-Net, integrating a cross-attention mechanism with ESM-2 and graph convolutional network (GCN), to identify candidate RhlA mutants. Via an approach combining deep learning-based activity prediction, Rosetta analysis, molecular docking, and molecular dynamics simulations, four novel RhlA mutants (R74A_L148C_S173K, R74A_A101M_S173T, R74A_S173L_Q176L, and R74A_L148C_S173A) were screened from a targeted mutant library. Structural analyses revealed that these mutants form stable conformations, enhancing substrate binding affinity. In wet-lab validation, the candidate mutants exhibited superior catalytic potential, with the enzymatic activity of R74A_L148C_S173A reaching 373.38 U/mg, representing a 3.6-fold increase compared to the wild-type enzyme. The remaining mutants also maintained high activity levels (290-317 U/mg). In summary, this study provides an EGCA-Net-based screening framework for the rapid identification and in-depth characterization of novel enzyme mutants.

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder. Adults with ADHD continue to exhibit deficits in attention and executive function, whereas the neural mechanisms underlying visual mismatch negativity (vMMN)-related processing remain unclear. This study investigated the functional brain connectivity characteristics of adults with ADHD by combining standardized low-resolution brain electromagnetic tomography (sLORETA), phase-locking value (PLV) analysis, and a graph convolutional network (GCN) based on a visual Oddball paradigm that elicited vMMN responses. Electroencephalography (EEG) data were collected from 10 adults with ADHD and 10 healthy controls using a 128-channel recording system. Source signals from 68 brain regions defined by the Desikan-Killiany atlas were reconstructed using sLORETA. Functional connectivity networks were constructed using PLV and subsequently classified by the GCN model. The results showed that the accuracy, precision, recall, and F1-score of the GCN model under five-fold cross-validation were (85.13 ± 1.94)%, (80.58 ± 2.08)%, (86.04 ± 1.76)%, and (83.21 ± 1.89)%, respectively. Node feature weights and classification contribution analyses identified the lingual gyrus, calcarine fissure and surrounding cortex, parahippocampal gyrus, and precuneus as highly discriminative brain regions. These findings indicate that adults with ADHD exhibit abnormal functional connectivity patterns during vMMN-related processing and provide evidence for the auxiliary identification and neural mechanism investigation of ADHD.

Aggregating global-scale pixel-wise forgery cues within a graph.

Using surfactants to manipulate the interfacial tension (IFT) of oil-water systems represent a critical strategy for enhanced oil recovery (EOR). However, predicting the physical properties of surfactants based on their molecular structure remains a challenging task, as conventional machine learning methods struggle to capture the coupled interactions between molecular structures and environmental parameters, while existing graph neural networks predominantly focus on single-molecule representations and overlook the characteristics of the system environment. Accordingly, a Gated Message-passing Graph Neural Network with an Attention Mechanism (Gated-MPNN-AT) is proposed to integrate molecular graph structures and environmental features, aiming to achieve accurate prediction of interfacial tension in surfactant-oil-water systems. The model dynamically controls the message passing process through a dual gated mechanism, adopts a Cross-Attention mechanism to achieve the in-depth interaction between molecular topological features and environmental parameters, and designs a hybrid robust loss function to handle the IFT data with cross-order-of-magnitude distribution. The research results show that the prediction accuracy of the model is better than that of traditional machine learning methods (such as Random Forest (RF) and eXtreme Gradient Boosting (XGBoost)) and some graph neural network methods (such as Graph Convolutional Network (GCN), and Graph Attention Network (GAT)). Ablation experiments have confirmed that the gated mechanism increases the coefficient of determination (R 2) by 4.8%, and the Cross-Attention fusion strategy reduces the mean absolute error (MAE) by 21.3%. Meanwhile, the model has good generalization ability and strong anti-interference ability against abnormal IFT data.